If you're not at the table, you're on the menu

AI-forward: Claude, an Anthropic model, drafted this from my notes, about Anthropic’s protocol, and checked every source against its claim. The argument, the cuts and the edit are mine.

In two years the Model Context Protocol went from one company’s open-source release to the way AI agents reach everything else. It won by being the first one people could use that afternoon, not by being the best.

What is MCP, actually?

A model needs tools to do anything, and each tool needs a connector. Before MCP, every AI app wrote its own connector for every tool, so ten apps and a hundred tools meant a thousand bits of glue.

MCP, which Anthropic released in November 2024, turns ten times a hundred into ten plus a hundred. A tool publishes one MCP server, and any app that speaks MCP can use it. Microsoft’s Language Server Protocol did the same for code editors in 2016, and the MCP specification says MCP takes some inspiration from it.

The timeline, before it was sacred

MCP
merged in
pruned
still open
Agentic AI Foundation

Every line here tried to let software act in the world for us. Most were pruned. The survivors now run into one trunk.

In Loki, the Time Variance Authority prunes branches off the sacred timeline. In protocols, use does.

In the early 1990s, agent researchers built a language for software agents to ask and tell each other things (KQML), which FIPA ACL later succeeded. Around 2000, web services tried to give machines a phone book: describe your service in WSDL, list it in UDDI. The big public UDDI registry closed in 2006. Both died, not because the idea was wrong, but because nobody needed it yet. Language models made the need real:

2022 to early 2023: research showed a model could reason and call tools in a loop (ReAct, then Toolformer).
March 2023: ChatGPT plugins. A website described its API in a manifest and ChatGPT could call it. OpenAI wound plugins down a year later.
From June 2023: function calling, from OpenAI, then Google and Anthropic, each in its own format.
April 2025: Google’s Agent2Agent protocol (A2A), for agents to hand work to each other. IBM’s Agent Communication Protocol merged into it that summer.

Then the branches folded in. OpenAI adopted MCP in March 2025, four months after launch. Google said Gemini would support it the next month. Microsoft has been building it into Windows as a public preview since late 2025. OpenAI’s apps platform for ChatGPT, launched that October, runs on MCP. In December 2025 Anthropic gave MCP to a new Linux Foundation body, the Agentic AI Foundation, alongside OpenAI’s AGENTS.md and Block’s goose. In August 2026, Google’s A2A moved under the same roof.

In Loki, one being imposed the sacred timeline. Handing MCP to a foundation is what stopped it being Anthropic’s, and it’s why the others joined.

Didn’t Google do the work?

A lot. A2A, with agents finding each other, negotiating and handing off long tasks, is the bigger idea. Google engineers co-proposed WebMCP, which lets an ordinary web page offer tools to an agent in the browser. Yet Google now runs managed MCP servers for its own cloud.

Google bet on a later problem. Every developer needed an agent connected to a tool that week. MCP solved that, with an SDK and a desktop app that used it on day one. A2A solved a problem most teams don’t have yet.

Is MCP any good?

Good enough, but the criticisms are fair:

Security came second. Proper authorisation arrived months after launch. The model trusts a tool’s description, so a malicious server can steer it. Give one agent private data, untrusted content and a way to send things out, and you’ve built a leak. Simon Willison calls it the lethal trifecta.
It eats the model’s attention. Every tool’s description fills the model’s working memory. Anthropic and Cloudflare have both published ways round this that have the model write code against MCP servers instead.
The first design was hard to scale. Long-lived connections suit a laptop, not a server farm. The July 2026 spec moved the core to plain request and response.

Richard Gabriel named this pattern in 1989: worse is better. The simple thing ships, spreads, then everyone patches it in place, because nobody rewires once the sockets are in every wall. MCP’s own docs compare it to USB-C.

Will it take over the web?

MCP is becoming the web’s second front door, the one agents use. If your product has no MCP server, an agent routes round you. Or someone wraps your API in an unofficial server and decides how agents see you. Either way you’re on the menu.

I’ve seen this in EV charging. Once an open protocol let chargers and back offices talk to each other, every vendor’s question stopped being “do we support it?” and became “what’s still ours?” The ones who helped write the protocol shaped the answer.

What to do

If you sell software, publish your own MCP server before someone else does, and describe your product in your own words. Keep it small, because nobody gets all the agent’s working memory. Review tool descriptions as product copy and as attack surface. If this matters to your business, get in the room where it’s decided: that’s the table.